def __init__(self, global_net, summary_writer, global_steps_counter, args):
self.env = make_env(args, record_video=args.record_video)
self.global_net = global_net
self.summary_writer = summary_writer
self.global_steps_counter = global_steps_counter
self.eval_every = args.eval_every
self.eval_times = 0
self.eval_episodes = args.eval_episodes
self.saver = tf.train.Saver(max_to_keep=5)
self.model_dir = os.path.join(args.save_path, 'models/')
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
def __init__(self, global_net, summary_writer, global_steps_counter, args):
self.env = make_env(args, record_video=args.record_video)
self.global_net = global_net
self.summary_writer = summary_writer
self.global_steps_counter = global_steps_counter
self.eval_every = args.eval_every
self.eval_times = 0
self.eval_episodes = args.eval_episodes
self.saver = tf.train.Saver(max_to_keep=5)
self.model_dir = os.path.join(args.save_path, 'models/')
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
def main(args):
if args.save_path is not None and not os.path.exists(args.save_path):
os.makedirs(args.save_path)
summary_writer = tf.summary.FileWriter(os.path.join(args.save_path, 'log'))
global_steps_counter = itertools.count() # thread-safe
global_net = Net(S_DIM, A_DIM, 'global', args)
num_workers = args.threads
workers = []
# create workers
for i in range(1, num_workers + 1):
worker_summary_writer = summary_writer if i == 0 else None
worker = Worker(i, make_env(args), global_steps_counter,
worker_summary_writer, args)
workers.append(worker)
saver = tf.train.Saver(max_to_keep=5)
with tf.Session() as sess:
coord = tf.train.Coordinator()
if args.model_path is not None:
print('Loading model...\n')
ckpt = tf.train.get_checkpoint_state(args.model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('Initializing a new model...\n')
sess.run(tf.global_variables_initializer())
print_params_nums()
# Start work process for each worker in a seperated thread
worker_threads = []
for worker in workers:
t = threading.Thread(target=lambda: worker.run(sess, coord, saver))
t.start()
time.sleep(0.5)
worker_threads.append(t)
if args.eval_every > 0:
evaluator = Evaluate(global_net, summary_writer,
global_steps_counter, args)
evaluate_thread = threading.Thread(
target=lambda: evaluator.run(sess, coord))
evaluate_thread.start()
coord.join(worker_threads)
def main(args):
if args.save_path is not None and not os.path.exists(args.save_path):
os.makedirs(args.save_path)
summary_writer = tf.summary.FileWriter(os.path.join(args.save_path, 'log'))
global_steps_counter = itertools.count() # thread-safe
global_net = Net(S_DIM, A_DIM, 'global', args)
num_workers = args.threads
workers = []
# create workers
for i in range(1, num_workers + 1):
worker_summary_writer = summary_writer if i == 0 else None
worker = Worker(i, make_env(args), global_steps_counter,
worker_summary_writer, args)
workers.append(worker)
saver = tf.train.Saver(max_to_keep=5)
with tf.Session() as sess:
coord = tf.train.Coordinator()
if args.model_path is not None:
print('Loading model...\n')
ckpt = tf.train.get_checkpoint_state(args.model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('Initializing a new model...\n')
sess.run(tf.global_variables_initializer())
print_params_nums()
# Start work process for each worker in a seperated thread
worker_threads = []
for worker in workers:
t = threading.Thread(target=lambda: worker.run(sess, coord, saver))
t.start()
time.sleep(0.5)
worker_threads.append(t)
if args.eval_every > 0:
evaluator = Evaluate(
global_net, summary_writer, global_steps_counter, args)
evaluate_thread = threading.Thread(
target=lambda: evaluator.run(sess, coord))
evaluate_thread.start()
coord.join(worker_threads)